RynnUnify

RynnUnify combines an 8B autoregressive reasoner with an 8B video generator for a total of 16B parameters. The reasoner predicts the next subtask, its visual description and duration, and the generator attends to the reasoner representations through joint attention to produce a video segment conditioned on scene memory and recent motion. Interleaved visual planning uses cross-chunk scene memory, self-forcing training and adaptive durations. An RL-trained agentic extension uses Branch-Conditional Policy Optimization so the model can answer directly or generate visual evidence before answering. The World-Action Model extension is trained jointly on video and actions for composite manipulation. The model scores 57.9 average on general embodied understanding benchmarks and improves the Rynn-VP overall visual planning score by 9.75 points over Cosmos 3 Nano.